From 78110352943f20bc6e0c9dfe77b169575bde899a Mon Sep 17 00:00:00 2001
From: Michael Schellenberger Costa <miscco@nvidia.com>
Date: Thu, 30 Jan 2025 09:38:44 +0100
Subject: [PATCH] Try to always include the definition of barrier_native_handle
 when needed (#3556)

---
 libcudacxx/include/cuda/__barrier/barrier_native_handle.h     | 4 ++++
 .../include/cuda/__memcpy_async/try_get_barrier_handle.h      | 1 +
 libcudacxx/include/cuda/barrier                               | 1 +
 3 files changed, 6 insertions(+)

diff --git a/libcudacxx/include/cuda/__barrier/barrier_native_handle.h b/libcudacxx/include/cuda/__barrier/barrier_native_handle.h
index 29879c71edf..a685c832723 100644
--- a/libcudacxx/include/cuda/__barrier/barrier_native_handle.h
+++ b/libcudacxx/include/cuda/__barrier/barrier_native_handle.h
@@ -25,6 +25,8 @@
 #include <cuda/__fwd/barrier.h>
 #include <cuda/std/cstdint>
 
+#if _CCCL_HAS_CUDA_COMPILER
+
 _LIBCUDACXX_BEGIN_NAMESPACE_CUDA_DEVICE
 
 _CCCL_DEVICE inline _CUDA_VSTD::uint64_t* barrier_native_handle(barrier<thread_scope_block>& __b)
@@ -34,4 +36,6 @@ _CCCL_DEVICE inline _CUDA_VSTD::uint64_t* barrier_native_handle(barrier<thread_s
 
 _LIBCUDACXX_END_NAMESPACE_CUDA_DEVICE
 
+#endif // _CCCL_HAS_CUDA_COMPILER
+
 #endif // _CUDA___BARRIER_BARRIER_NATIVE_HANDLE_H
diff --git a/libcudacxx/include/cuda/__memcpy_async/try_get_barrier_handle.h b/libcudacxx/include/cuda/__memcpy_async/try_get_barrier_handle.h
index d2207faf91d..cb55695031b 100644
--- a/libcudacxx/include/cuda/__memcpy_async/try_get_barrier_handle.h
+++ b/libcudacxx/include/cuda/__memcpy_async/try_get_barrier_handle.h
@@ -22,6 +22,7 @@
 #endif // no system header
 
 #include <cuda/__barrier/barrier_block_scope.h>
+#include <cuda/__barrier/barrier_native_handle.h>
 #include <cuda/std/__atomic/scopes.h>
 #include <cuda/std/__barrier/barrier.h>
 #include <cuda/std/__barrier/empty_completion.h>
diff --git a/libcudacxx/include/cuda/barrier b/libcudacxx/include/cuda/barrier
index 0d65d4bf344..bcc2999cb31 100644
--- a/libcudacxx/include/cuda/barrier
+++ b/libcudacxx/include/cuda/barrier
@@ -34,6 +34,7 @@
 #include <cuda/__barrier/barrier_arrive_tx.h>
 #include <cuda/__barrier/barrier_block_scope.h>
 #include <cuda/__barrier/barrier_expect_tx.h>
+#include <cuda/__barrier/barrier_native_handle.h>
 #include <cuda/__barrier/barrier_thread_scope.h>
 #include <cuda/__memcpy_async/memcpy_async.h>
 #include <cuda/__memcpy_async/memcpy_async_tx.h>